Market Equilibrium under Piecewise Leontief Concave Utilities

Market Equilibrium under Piecewise Leontief
Concave Utilities
Jugal Garg∗
Max-Planck-Institut f¨
ur Informatik, Germany
[email protected]
Abstract. Leontief function is one of the most widely used function
in economic modeling, for both production and preferences. However it
lacks the desirable property of diminishing returns. In this paper, we
consider piecewise Leontief concave (p-Leontief) utility function which
consists of a set of Leontief-type segments with decreasing returns and
upper limits on the utility. Leontief is a special case when there is exactly
one segment with no upper limit.
We show that computing an equilibrium in a Fisher market with
p-Leontief utilities, even with two segments, is PPAD-hard via a reduction from Arrow-Debreu market with Leontief utilities. However, under a
special case when coefficients on segments are uniformly scaled versions
of each other, we show that all equilibria can be computed in polynomial time. This also gives a non-trivial class of Arrow-Debreu Leontief
markets solvable in polynomial time.
Further, we extend the results of [22,5] for Leontief to p-Leontief
utilities. We show that equilibria in case of pairing economy with pLeontief utilities are rational and we give an algorithm to find one using
the Lemke-Howson scheme.
1
Introduction
Market equilibrium is a fundamental concept in mathematical economics and has
been studied extensively since the work of Walras [20]. The notion of equilibrium
is inherently algorithmic, with many applications in policy analysis and recently
in e-commerce [9,12,18]. The Arrow-Debreu (exchange) market model consists of
a set of agents and a set of goods, where each agent has an initial endowment of
goods and a utility (preference) function over bundle of goods. At equilibrium,
each agent buys a utility maximizing bundle from the money obtained by selling
its initial endowment and the market clears.
It is customary in economics to assume utility functions to be concave and
satisfying the law of diminishing returns. Leontief utility function is a wellstudied concave function, where goods are complementary and they are needed
in a fixed proportion for deriving a positive utility, for e.g., bread and butter.
It is a homogeneous function of degree one, where the utility is multiplied by α
∗
Work done while at Georgia Tech, Atlanta, USA.
when the amount of each good is multiplied by α, for any α > 0; hence it does
not, as such, model diminishing returns. Consider the following example.
Example. Suppose Alice wants to consume sandwiches and for making a sandwich, she needs two slices of bread and one slice of cheese. It seems her utility
function for bread and cheese can be modeled as a Leontief function, but it is
not appropriate because her utility for the second sandwich is less than the first
one due to satiation, and so on.
In this paper, we define piecewise Leontief concave (p-Leontief) utility function,
which not only generalizes Leontief but also captures diminishing returns to
scale and seems to be more relevant in economics. Further, we derive algorithmic
and hardness results for both Fisher∗ and exchange market models under these
functions. A p-Leontief utility function consists of a set of segments, where utility
obtained on each segment is as per a Leontief function with a limit on the utility.
The extra bundle needed, to obtain another unit of utility on segment k, is
strictly more than that on segment k–1† . This puts a natural ordering on the
segments, so that function remains concave and captures diminishing returns
(see Section 2.4 for the precise definition). Observe that a Leontief function is
simply a p-Leontief function with exactly one segment and no upper limit.
Recall the above example. Alice’s utility for bread and cheese can be modeled
as a p-Leontief function as follows: On the first segment, a unit of utility can be
derived by consuming 2 slices of bread and 1 slice of cheese, and the upper limit
is 1, i.e., at most 1 unit of utility can be derived on this segment. On the second
segment, a unit of utility can be derived by consuming 4 slices of bread and 2
slices of cheese, and the upper limit is 2, and so on.
Since Leontief function is a special case of p-Leontief function, all hardness
results for Leontief utilities [5] simply carry over to p-Leontief utilities and we
get the following theorem.
Theorem 1 (Hardness of Exchange p-Leontief ). Computing an equilibrium in an exchange market with p-Leontief utilities is PPAD-hard, and all
equilibria can be irrational even if all input parameters are rational numbers.
There is a qualitative difference between the complexity of computing an
equilibrium in Fisher and Arrow-Debreu markets under Leontief utilities; while
polynomial time in the former case through Eisenberg’s convex program [11], it
is PPAD-hard in the latter case [5]. In contrast, we show that Fisher is no easier
than Arrow-Debreu under p-Leontief utilities and obtain the following theorem.
Theorem 2 (Hardness of Fisher p-Leontief ). Computing an equilibrium in
Fisher market with p-Leontief utilities, even with two segments, is PPAD-hard.
For the above theorem, we essentially give a reduction from exchange pLeontief with k segments to Fisher p-Leontief with k + 1 segments. Further we
∗
†
A special case of exchange market model, defined in Section 2.2.
By strictly more than, we mean at least one good is needed in greater amount.
show that when coefficients on segments are uniformly scaled versions of each
other, then Fisher market equilibria can be computed in polynomial time. This
special case arises in many practical situations, like in the above example of
Alice’s utility function for bread and cheese. The proportion of bread and cheese
on each segment remains 2:1, however her utility per unit of sandwich decreases.
This also gives us a non-trivial class of tractable exchange Leontief markets,
where the sum of endowment matrix (W ) and Leontief utility coefficient matrix
(U ) is a constant times all one matrix (see Section 4.1 for details). We note
that Fisher is a special case of exchange market model for which W is very
special, however an exchange market satisfying our condition does not require
W to be special and hence it does not arise from a Fisher market. To the best
of our knowledge, apart from the Fisher markets, we are not aware of any other
non-trivial tractable classes of exchange Leontief markets.
Pairing economy is a special case of exchange markets, where each agent
brings a different good to the market (see Section 2.1 for precise definition). In
case of a pairing economy with Leontief utilities, [5] showed that equilibria are
rational and they are in one-to-one correspondence with the symmetric Nash
equilibria in a symmetric bimatrix game. We extend the results of [5,22] for
Leontief to p-Leontief and obtain the following (informal) theorem.
Theorem 3 (Pairing Economy: Rationality and Algorithm). In a pairing
economy with p-Leontief utilities, equilibrium prices are rational if all input parameters are rational. Further computing an equilibrium is PPAD-complete and
there is a finite time algorithm to find one using the Lemke-Howson scheme.
For this, we first characterize equilibrium conditions for exchange market
with p-Leontief utilities using the right set of variables: a variable to capture
price for each good and a variable to capture utility on each segment. In case of
pairing economy, these conditions can be divided into two parts. The first part
captures the utility on each segment, at equilibrium, as a linear complementarity problem (LCP) formulation, where we use the power of complementarity to
ensure that segments are allocated in the correct order. The second part is a
linear system of equations of type Ap = p in prices given the utilities on each
segment. Next we show that the LCP of first part can be solved using the classic
Lemke-Howson scheme [14] and then we obtain the equilibrium prices by solving
Ap = p. For this to work, we need a positive solution of Ap = p, which is guaranteed by the Perron-Frobenius theorem because A turns out to be a positive
stochastic matrix.
Related work. Since Leontief is a homogeneous function of degree one, equilibria in a Fisher market with Leontief utilities are captured by Eisenberg’s convex
program [11] and hence it is computable in polynomial time. Mas-Colell (in [10]
by Eaves) gave an example of Leontief economy (both Fisher and exchange),
where all equilibria are irrational, even if the input parameters are rational numbers; this discards the possibility of LCP based approach for Leontief economy.
Pairing economy model is used by [5] to show that computing an equilibrium
in an exchange market with Leontief utilities is PPAD-hard and it has also been
studied in many other settings, for e.g., [22,21,7,23].
There are generalizations of Leontief studied by [22,4]. [22] considered a class
of piecewise linear concave (PLC) utility functions and showed that equilibrium
in pairing economy is equivalent to solving an LCP. [4] studied market equilibrium under hybrid linear-Leontief utility function. Leontief is a special case in
both of them, however these classes are still homogeneous of degree one and do
not model the diminishing returns to scale. Hence they are quite different and
not comparable with p-Leontief.
2
Preliminaries
2.1
Exchange Market
Exchange is a most fundamental market model and it is extensively studied since
the work of Walras [20]. An exchange market consists of a set of agents A and a
def
def
set of goods G. Let m = |A| and n = |G|. Each agent comes to the market with
an initial endowment of goods, where Wij is the amount of good j with agent i,
and a utility function ui : Rn+ → R+ over bundle of goods. Given prices
P of goods
p = (p1 , . . . , pn ), where pj is the price of good j, each agent i earns j∈G Wij pj
by selling its initial endowment and buys a (optimal) bundle which maximizes
its utility function from the earned money. At equilibrium prices, market clears,
i.e., demand of each good matches with its supply.
It is customary in economics to assume utility functions to be non-negative,
non-decreasing and concave; non-negative and non-decreasing because of freedisposal property, and concave to model the diminishing returns. The celebrated
Arrow-Debreu theorem [1] shows that market equilibrium exists for a very general class of utility functions under some mild conditions. Further, we note that
equilibrium prices in this case is scale invariant, i.e., if p is equilibrium prices,
then so is αp, ∀α > 0. And, it is without loss
P of generality, to assume that the
total initial endowment of each good, i.e., i Wij = 1, ∀j ∈ G, is unit‡ .
Pairing Economy. In pairing economy, the initial endowment of each agent is
a good which is different than the goods brought by other agents – agents and
goods are paired up. In this case, W is an identity matrix, i.e., W = I.
2.2
Fisher Market
Fisher market model was defined by Irving Fisher in 1891 [2], where unlike
exchange market, buyers and sellers are two different entities. A Fisher market
consists of a set of agents A and a set of goods G. Each agent i has money
Ei , and a utility function ui : Rn+ → R+ over bundle of goods. Let Qj denotes
‡
This is like redefining the unit of goods by appropriately scaling utility parameters.
the quantity of good j in the market. Given prices of goods p = (p1 , . . . , pn ),
where pj is the price of good j, each agent i buys a (optimal) bundle which
maximizes its utility function subject to its budget constraints. At equilibrium
prices, market also clears, i.e., demand of each good matches with its supply.
It is a special case of exchange market: set Wij = Qj PEiEi , ∀(i, j), and keep
i
the ui ’s unchanged.
2.3
Leontief Utility Function
Leontief is a well studied utility function in economics, where goods are complementary and they are needed in a fixed proportion for deriving positive utility,
for e.g. bread and butter. Formally, from a bundle xi = (xi1 , . . . , xin ) of goods,
a Leontief utility function of agent i may be defined as
ui (xi ) = min{
j
xij
}.
Uij
Here agent i derives one unit of utility when it gets Uij amount of each good
j. If the utility function of each agent is Leontief, then it can be represented
as a matrix U = [Uij ]i∈A,j∈G , whose ith row Ui = [Uij ]j∈G contains all the
coefficients of agent i.
A way to represent Leontief utility function is by choosing x-axis to denote the
amount of a good whose coefficient is positive. The amount of the remaining goods
are just a fixed proportion of this amount.
Fig. 1 depicts a simple example of Leontief
utility function on two goods. The goods
are required in 2:1 ratio, i.e., one unit of
utility is obtained from consuming 2 units
of good 1 and 1 unit of good 2.
Utility
2:1
1
2
Amount of good 1
Fig. 1. Leontief utility function
Leontief is a special class of Constant Elasticity of Substitution (CES) [15]
utility functions; when CES parameter approaches −∞ we get Leontief. We note
that all CES functions are homogeneous of degree one, and so is Leontief, i.e.,
ui (αxi ) = αui (xi ), ∀α > 0. In other words, scaling a bundle by α > 0 scales the
utility by the same factor. Thus it does not, as such, capture diminishing returns,
an important property to model real-life preferences like the example of Alice’s
utility function in the introduction. To circumvent this, in the next section, we
define piecewise-Leontief concave utility function, which extends Leontief and
allows to model diminishing returns in a logical manner.
2.4
p-Leontief Utility Function
In this section, we define piecewise Leontief concave (p-Leontief) utility function
by extending Leontief to capture diminishing returns to scale. Such a function
consists of a set of segments (pieces), where the utility derived on each segment
is a Leontief function with an upper limit. Further, the coefficient of segments
are such that the function remains concave.
Formally, let ui : Rn+ → R+ be the p-Leontief utility function of agent i with
l segments. Since each segment k represents a Leontief function with a limit, it
k
]j∈G , Lki ), where Uik stores the coefficients of
can be represented as (Uik = [Uij
Leontief function and Lki stores the limit. The utility ui from a bundle xi =
(xi1 , . . . , xin ) is defined as
(
ui (xi ) = min
j
xij 1
1 , Li
Uij
)
(
(
+ max min
j
1
xij − L1i Uij
, L2i
2
Uij
(
(
· · · + max min
xij −
j
A way to represent p-Leontief utility function is by choosing x-axis to denote the
amount of a good whose coefficient is positive. The amount of the remaining goods
can be easily obtained from this amount.
Fig. 2 depicts a simple example of p-Leontief
utility function from two goods. It has three
segments; on the first segment, one unit of
utility is obtained from 2 units of good 1 and
1 unit of good 2, and the maximum utility
that can be derived at this rate is L1 . After
that the rate decreases to one unit of utility
from 4 units of good 1 and 2 units of good
2 on the second segment, and the maximum
utility at this rate is L2 , and so on.
)
Pl−1
k=1
l
Uij
)
,0
k
Lki Uij
+ ···
)
, Lli
)
,0 .
Utility
5:3
L2
4:2
L1
2:1
Amount of good 1
Fig. 2. p-Leontief utility function
k
Note that Lki Uij
is the amount of good j needed to utilize segment k completely; essentially segments are utilized in order from 1 to l. Further, there is
no upper limit on the last segment, i.e., Lli = ∞. To ensure diminishing returns
and concavity, we enforce Uik+1 ≥ Uik , ∀k ≥ 1.
Next we show that such a function is indeed concave. For this, consider
the following linear program (LP) to compute the utility from a bundle xi =
(xi1 , . . . , xin ), where βik captures the utility obtained on segment k, and xkij
captures the amount of good j allocated on segment k.
max
l
X
βik
k=1
l
X
xkij = xij ,
∀j ∈ G
(LP)
k=1
βik ≤
xkij
k
Uij
, 1 ≤ k ≤ l, ∀j ∈ G
0 ≤ βik ≤ Lki ,
1 ≤ k ≤ l.
Lemma 4. For any given xi , optimal solution of (LP) gives ui (xi ).
Proof. Since Uik+1 ≥ Uik , ∀k, it is easy to check that optimal solution of (LP) will
allocate segments in an increasing order, i.e., βik+1 > 0 ⇒ βik = Lki , ∀k. Further
k
βik Uij
= xkij and it implies that βik = max{minj {
xij −
Pk−1
a=1
k
Uij
a
La
i Uij
, Lki }, 0}. This
t
u
proves the claim.
Lemma 5. p-Leontief is a concave function.
e i , we have ui (λxi +
Proof. We need to show that for any two bundles xi and x
k
ek , x
(1 − λ)e
xi ) ≥ λui (xi ) + (1 − λ)ui (e
xi ), ∀λ ∈ (0, 1). Let (βi , xkij ) and (β
ek ) be
Pi ij
e i respectively, so ui (xi ) = k βik and
the optimal solutions of (LP) for xi and x
P ek
ui (e
xi ) =
β using Lemma 4.
i
k
k
Let z i = λxi + (1 − λ)e
xi . Consider a candidate zij
= λxkij + (1 − λ)e
xkij and
k
k
k
x
x
e
z
e k . Further we have
γik = minj { Uijk } = minj {λ Uijk + (1 − λ) Uijk } ≥ λβik + (1 − λ)β
i
ij
ij
ij
k
e ≤ Lk . It is easy to check that (γ k , z k ) satisfies all the
γik ≤ Lki as both βik , β
i
i
i
ij
P k
P k
P ek
constraints of (LP), and the objective value is k γi ≥ λ k βi +(1−λ) k β
i.
Therefore, optimal solution of (LP) at z i will be at least λui (xi ) + (1 − λ)ui (e
xi ),
which proves the claim.
t
u
3
Exchange Market with p-Leontief Utility Functions
In this section, we study exchange markets with p-Leontief utilities. First we
show that the problem of computing an equilibrium is hard, and then design
a finite time algorithm to compute an equilibrium through an linear complementarity problem (LCP) based approach. Since Leontief is a special case of
p-Leontief, the next theorem follows from [5,10].
Theorem 6. – Checking existence of an equilibrium in p-Leontief exchange
markets is NP-hard.
– Assuming sufficiency condition§ , computing an equilibrium is PPAD-hard.
– All equilibrium prices can be irrational even if all input parameters are rational numbers.
For pairing economy with Leontief utilities, where each agent brings a different good to the market, [5,22] showed that there exists a rational equilibrium (if
one exists), and under a sufficiency condition (see Section 3.1), they are in oneto-one correspondence with symmetric Nash equilibria in a symmetric bimatrix
game. Thus equilibrium computation problem in this case is PPAD-complete.
We extend all these results to pairing economy with p-Leontief utilities. We
show that there is a rational equilibrium (if one exists) and under the similar
sufficiency condition, the problem of finding one is PPAD-complete and it can
be obtained using classic Lemke-Howson algorithm [14], which is known to run
fast in practice.
Next we begin with the characterization of market equilibrium for pairing
economy with p-Leontief utilities.
3.1
Market Equilibrium Characterization
Recall that at market equilibrium, each agent obtains a utility maximizing (optimal) bundle of goods subject to its budget constraints, and market clears. In
case of pairing economy, each agent i brings one unit of good i to the market
and therefore its budget is the price assigned to its good. Since utility function
on each segment is Leontief, goods are consumed in a fixed proportion on a
k
segment, determined by Uij
’s, at an optimal bundle.
Let βik denote the utility obtained on segment k by agent i, and xkij denote
the amount of good j consumed on segment k by agent i. Note that at least
k
βik Uij
amount of each good j is needed in deriving βik utility. Therefore, at an
k k
optimal bundle, we have xkij = Uij
βi , ∀(k, i, j). Let pj denote the price of good
k
j. Since xij ’s can be obtained from βik ’s, the real set of variables to capture are
β = [βik ]i∈A,k∈[l] and p = [pj ]j∈G . Market clearing constraints for each good
are:
X
X
k k
xkij =
Uij
βi ≤ 1, ∀j ∈ G
(1)
i,k
i,k
Similarly, market clearing constraints for each agent are:
X
X
k k
xkij pj =
Uij
βi pj = pi , ∀i ∈ A
j,k
(2)
j,k
P
k k
From (1) and (2), we can conclude that if pj > 0 then i,k Uij
βi = 1. This
can be written as the following complementarity constraint:
X
X
k k
k k
Uij
βi ≤ 1; pj ≥ 0; pj (
Uij
βi − 1) = 0, ∀j ∈ G
i,k
§
refer to Section 3.1
i,k
Further, we need that if βik+1 > 0 then βik = Lki , which can be captured as
the following complementarity constraint:
βik+1 ≥ 0;
βik ≤ Lki ;
βik+1 (βik − Lki ) = 0,
∀(k, i)
We avoid all-zeros solution by putting the following constraint:
X
l
Uij
pj > 0, ∀i ∈ A
j
Putting these constraints together, next consider the following conditions in
variables (β, p)
∀j ∈ G :
X
k k
Uij
βi ≤ 1;
pj ≥ 0;
i,k
∀i ∈ A :
X
X
k k
pj (
Uij
βi − 1) = 0
(3.1)
i,k
k k
Uij
βi pj = pi ;
βi1 ≥ 0
(3.2)
j,k
βik ≤ Lki ;
∀(k, i) :
∀i ∈ A :
X
l
Uij
pj > 0
βik+1 ≥ 0;
βik+1 (βik − Lki ) = 0
(3.3)
(3.4)
j
Remark 7. Note that the equality in (3.2) is quadratic, which eventually leads
to irrational equilibria.
Lemma 8 (Equilibrium Characterization). (β, p) gives a market equilibrium of pairing economy iff it satisfies (3.1)-(3.4).
Proof. It is easy to see that market equilibrium (x, p) gives a solution of (3.1)(3.4), where βik =
xk
ij
k .
Uij
Market clearing ensures (3.1) and (3.2) and allocating
P k
P k+1
segments in order ensures (3.3). Since j Uij
pj ≤ j Uij
pj , ∀k, if (3.4) is not
satisfied for some agent i, then it would have demanded infinite amount of some
goods, contradicting market clearing.
k k
For the other direction, let (β, p) be a solution of (3.1)−(3.4), set xkij = Uij
βi .
Then clearly, (3.1) ensures that no good is sold more than its supply and if a
good is under-sold then its price is zero. Further (3.2) ensures that agents do
not spend more than their earnings.
Now we only need to show that it gives an optimal bundle to each agent.
At an optimal bundle, goods are bought in a fixed proportion on each segment
k k
which follows from the fact that xkij = Uij
βi , and before allocating any amount
to segment k + 1, segment k should be utilized completely, which is ensured by
(3.3).
Finally (3.4) avoids the all-zeros solution, which does not give an equilibrium.
t
u
Remark 9. If we do not put (3.4), then it will capture quasi-equilibrium, where
we are not required to assign optimal bundles to zero income agents. Further,
(3.1)−(3.4) are general
P enough to capture equilibrium of exchange market when
pi is replaced with j Wij pj in (3.2).
In general, market equilibrium may not exist and checking existence is NPhard (Theorem 6). Next we show that under a sufficiency condition, equilibrium
exists and the problem of computing one is PPAD-complete and we give a finite
time algorithm using classic Lemke-Howson scheme for finding one. We also note
that it is unlikely to strengthen this, without sufficiency condition, because that
will imply NP=co-NP by a result of Megiddo [17].
Sufficiency Conditions Arrow-Debreu [1] gave the following sufficiency conditions for the existence of market equilibrium: (i) W > 0, and (ii) each agent
is non-satiated, i.e., for each bundle, there is another bundle giving a better
utility. Clearly, p-Leontief utility function satisfy non-satiation, however under
pairing economy we have W = I and therefore condition (i) is not satisfied. In
fact, the following simple example, given in [5], shows that there may not be an
equilibrium, in general, for the pairing economy with Leontief utilities.


1 0 0
U = 1 1 0 
0 2 1
Further the sufficiency conditions, based on economy graph given in [16],
are not applicable for Leontief utility functions. Therefore, [5] assumed Uij >
0, ∀(i, j), and showed existence by reducing the problem to symmetric games.
Similarly, we assume
1
Uij
> 0, ∀(i, j)
(4)
and under this condition we design an algorithm to compute an equilibrium
for pairing economy with p-Leontief utilities, thereby also giving a constructive
proof of existence.
3.2
Algorithm
We assume that the conditions of (4) is satisfied for the rest of this section. Note
k
that (4) implies that Uij
> 0, ∀(i, j, k). Under the sufficiency condition, (3.4) is
automatically satisfied at a non-zero equilibrium prices, hence it is not needed.
Next we show that βi1 and pi are intimately related.
Lemma 10. At equilibrium βi1 > 0 iff pi > 0.
Proof. Lemma 8 implies that equilibrium satisfies (3.1)−(3.4). If pi > 0 then
agent i has money to buy a bundle which can give it positive utility, and hence
at equilibrium it will buy something on the first segment and that will imply
that βi1 > 0. For the other direction, if pi = 0, then agent i has no money
P k
P
pj = pi . Since p 6= 0 at
to buy anything. Further (3.2) implies k βik j Uij
P k
equilibrium, condition (4) ensures j Uij
pj > 0, ∀k, implying that βik = 0, ∀k,
t
u
and in particular βi1 = 0.
Using Lemma 10, and the fact that (3.4) is vacuously satisfied if p 6= 0, we
partition (3.1)−(3.3) into two parts as follows (like in [5] for Leontief):
X
∀j ∈ G :
LCP:
k k
Uij
βi ≤ 1;
βj1 ≥ 0;
βj1 (
i,k
∀(k, i) :
βik
∀i ∈ A :
X
k k
Uij
βi − 1) = 0
(5)
i,k
≤
Lki ;
X
βik+1
≥ 0;
k k
Uij
βi p j = p i ;
βik+1 (βik
pi ≥ 0;
−
Lki )
=0
p 6= 0
(6)
j,k
The next lemma follows using Lemma 8 and by the construction.
Lemma 11. A (β, p) is an equilibrium of a pairing economy with p-Leontief
utility function satisfying (4) iff it satisfies (5) and (6).
Due to Lemma 11 now our goal is to find a (β, p) satisfying both (5) and
(6). Note that (5) is a linear complementarity problem (LCP) in variables β.
Further, this LCP has a dummy solution β = 0 which does not correspond to
any solution of (6), hence market equilibrium. Using this, consider the algorithm
of Table 1 for finding an equilibrium.
Table 1. Algorithm
1. Solve LCP (5) using Lemke-Howson algorithm starting from β = 0,
and obtain a non-zero β.
2. Set pi = 0 if βi1 = 0
3. Obtain remaining prices using the first equality in (6)
The algorithm in Table 1 works only if the third step gives a positive solution.
The next theorem shows that this algorithm indeed gives an equilibrium.
Theorem 12 (Correctness). Assuming (4), the algorithm in Table 1 computes
a market equilibrium of a pairing economy with p-Leontief utility function.
Proof. First observe that the LCP in (5) is of type {Ay ≤ 1; y ≥ 0; y T (Ay −
1) = 0}, which is same as the LCP of finding a symmetric Nash equilibrium in
symmetric bimatrix game given by (A, AT ) [5]. Since Lemke-Howson algorithm
can find a non-zero y for such an LCP, we conclude that the first step of algorithm
in Table 1 gives us a non-zero β satisfying (5).
Next we set pi = 0Pwhenever βi1 = 0 and the remaining set of pj ’s is obtained
k k
through the equality j,k Uij
βi pj = pi , and it is a system of linear equalities of
1
type Cσ = σ. Under sufficiency condition Uij
> 0, ∀(i, j), C > 0 and stochastic,
i.e., sum of every column is 1 due to the complementarity condition of the first
condition in (5). By Perron-Frobenius theorem, C has a right positive eigenvector, i.e., σ > 0, which implies that we get all the remaining prices positive in
the third step of the algorithm. The solution (β, p) obtained from the algorithm
clearly satisfy both (5) and (6), hence it is a market equilibrium (Lemma 11). t
u
A number of results follows as corollaries using Theorem 12 and Lemma 11.
Corollary 13 (Existence). There exists an equilibrium in a pairing economy
with p-Leontief utility function satisfying (4).
Corollary 14 (Rationality). Equilibrium prices of pairing economy with pLeontief utility function are always rational if all input parameters are rational.
Proof. All the solutions of an LCP are rational when all input parameters are
rational [6]. Therefore, solutions of (5) are rational. Further, given a non-zero
solution β of (5), the third step of the algorithm 1 solves a system of linear equations with rational coefficients and hence computes rational prices that satisfies
(6). Thus corollary follows using Lemma 11.
t
u
Theorem 6 shows that the problem of computing an equilibrium is PPADhard. Using the rationality and the algorithm next we show it is also in PPAD.
Corollary 15 (PPAD-completeness). The problem of computing an equilibrium in pairing economy with p-Leontief utilities satisfying (4) is PPADcomplete.
Proof. Since the problem essentially reduces to finding a symmetric Nash equilibrium in a symmetric game, it is in PPAD. Further, we also have a reverse
reduction from game to market, as given in [5] for Leontief utilities, which makes
it PPAD-hard.
t
u
In summary, we obtained a finite time algorithm based on the Lemke-Howson
scheme to compute an equilibrium in pairing economy with p-Leontief utility function, which runs fast in practice in spite of the problem being PPADcomplete, and thereby extended the results of [5] to significantly general class of
utility functions.
4
Fisher Market with p-Leontief Utility Functions
We note that computing an equilibrium in a Fisher market with Leontief utilities
is equivalent to solving the Eisenberg’s convex optimization problem [11], and
hence it is polynomial time computable. In this section we show that this positive
result does not extend to markets with p-Leontief utility functions, even if all
the utility functions have only two segments. Essentially, we show that Fisher is
no easier than exchange under p-Leontief, which contrasts with the fact that the
behavior of exchange and Fisher under Leontief is entirely different – former is
PPAD-hard and latter is polynomial time.
Next we give a reduction from exchange market with Leontief utilities, where
U > 0 (see Section 2.3 for Leontief utilities), to Fisher market with p-Leontief
utility function which has two segments.
Theorem 16 (Reduction from Exchange to Fisher). An exchange market
M with Leontief utility functions, where U > 0, can be reduced to a Fisher market
M0 with p-Leontief utility function such that equilibria of M are in one-to-one
correspondence with equilibria of M0 (up to scaling).
Proof. Let M is defined by (W, U ) (see Section 2 for details), where W is initial
endowment matrix and U is Leontief utility function matrix, such that Uij >
0, ∀(i, j). WithoutPloss of generality, we assume that the total quantity of each
good is unit, i.e., i Wij = 1, ∀j. We will construct a Fisher market M0 defined
e 1, U
e 2 , L1 ), where E is initial money vector, Q is the quantity vector,
by (E, Q, U
k
e
U is Leontief utility function matrix for segment k for k = 1, 2, and L1 =
[L1i ]i∈A is the bound on maximum utility that can be obtained on the first
segment for each agent i. Note that there is no bound on the maximum utility
on the last segment.
e 1 , such that at any prices p,
The idea is to define the first segment, i.e., U
each agent i buys it completely and the remaining money after that is equal to
P
j Wij pj . Further, we also make sure that exactly unit amount of each good
remains after consuming the first segment of each agent entirely. In that case, it
e 2 = U , then on the second segment, Fisher becomes
is clear that if we define U
exactly same as exchange, and any equilibrium of M0 will give an equilibrium
of M and vice versa.
def
def
def
Let m = |A|, Umax = maxi,j Uij , and ∆ = mUmax . We set
Ei = 1,
∀i ∈ A
Qj = m + 1,
∀j ∈ G
1 m+1
1
eij
− Wij ), ∀(i, j) ∈ (A, G)
U
= (
∆ m
L1i = ∆,
∀i ∈ A
2
e = U.
U
e 1 > 0, ∀(i, j) as Wij ≤ 1, ∀(i, j), and U
e1 ≤ U
e 2 , ∀(i, j) due to the
Note that U
ij
ij
ij
P
P
choice of ∆.
condition implies that (m + 1) j pj = i Ei = m,
P Market clearing
m
we have
money spent on the first segment by agent i is
j pj = m+1 . The P
P 1 m+1
∆ P j ∆ ( m − Wij )pj = 1 − j Wij pj . The remaining money left with agent i
is j Wij pj to spend on the second segment, which is exactly equal to the money
with agent i in M. Further, the
Ptotal amount of good j needed to completely buy
first segment of each agent is i ( m+1
m − Wij ) = m, hence the remaining amount
e 2 is same as U , hence each
is 1, ∀j. Since the utility on the second segment U
0
market equilibrium of M will give a market equilibrium of M (up to scaling)
and vice versa.
t
u
Remark 17. (a) The above reduction in fact does not require U > 0, but without
e will not be concave and hence also not p-Leontief.
that, U
(b) The reduction can be easily extended to show that exchange market with
p-Leontief utility function, which has k segments reduces to Fisher market with
p-Leontief, which has k + 1 segments.
(c) The idea of designing first segment so that it consumes extra money was
also used in [19] for proving that Fisher market with separable piecewise-linear
and concave utilities is PPAD-hard. However our reduction is more general and
does not require the instance of exchange market to be special satisfying price
regulation property as in [19].
Corollary 18 (Hardness of Fisher p-Leontief ). Computing an equilibrium
in a Fisher market with p-Leontief utilities, even with two segments, is PPADhard.
Proof. We note that finding a symmetric Nash equilibrium in a two player symmetric game (A, AT ) is PPAD-hard [3,8]. Without loss of generality, we can
assume that A > 0 since adding the same constant to all entries does not change
the set of Nash equilibria. Further, [5] gave a reduction from the problem of
finding a symmetric Nash equilibrium in a symmetric game (A, AT ) to finding a
market equilibrium in exchange market with Leontief utilities (W = I, U = A),
where I is an identity matrix. Now the claim follows from Theorem 16.
t
u
4.1
A Special Case
In this section, we consider a special case of p-Leontief utility functions, where
the coefficients on segments are uniformly scaled versions of each other. Recall
that a Leontief utility function of an agent from a bundle x of goods is given
x
by minj∈G Ujj , where a unit of utility is obtained by consuming Uj amount of
each good j. A p-Leontief utility function has a set of Leontief-type segments
with upper limits on the utility. Segment k can be described by coefficients
U k = [Ujk ]j∈G . In general, Ujk ’s can be arbitrary non-negative numbers satisfying
U k ≤ U k+1 . However when they satisfy
Ujk11
Ujk21
=
Ujk12
Ujk22
, ∀(k1 , k2 ) and ∀(j1 , j2 ),
then we say that coefficients on segments are uniformly scaled versions of each
other, however the amount of each good j needed on a segment k + 1 for a unit
of utility is strictly greater than the amount needed on segment k.
This is an interesting special case and seems applicable in many practical
situations, like in the example of Alice’s utility function for bread and cheese in
the introduction.
Lemma 19. For the special case of p-Leontief utilities, where coefficients on
segments are uniformly scaled versions of each other, Fisher market equilibrium
can be computed in polynomial time.
Proof. We note that the Fisher market equilibrium with Leontief utilities can
be computed in polynomial time using the Eisenberg’s convex program. Now
observe that Fisher market equilibrium for Leontief utilities whose coefficients
are as per the first segments of p-Leontief utilities, also gives an equilibrium
for the special case as well. Here equilibrium prices and allocation map to the
exact same values, however the value of utility obtained at equilibrium can be
different.
t
u
Non-trivial tractable class of exchange Leontief markets. An exchange
market with Leontief utilities is described by an endowment matrix W and a
Leontief utility coefficient matrix U . Recall the reduction from Section 4. Using this reduction and Lemma 19, we can solve exchange market with Leontief
Utilities in polynomial time, when
1 m+1
(
− Wij ) = cUij , ∀(i, j),
∆ m
for a constant c < 1¶ . By redefining the units of goods, we can simplify this
condition to
W + U = cJ,
(7)
where c is a constant and J is all one matrix. This gives us a non-trivial class
of tractable exchange markets. We note that Fisher is a special case of exchange
market model for which W is very special, however an exchange market satisfying
(7) does not require W to be special and hence it does not arise from a Fisher
market. To the best of our knowledge, apart from the Fisher markets, we are not
aware of any other non-trivial tractable classes of exchange Leontief markets.
5
Discussion
We obtain both positive and negative results for p-Leontief utility function;
it is as easy as Leontief in case of pairing economy, however Fisher becomes
PPAD-hard. An interesting open question is whether exchange market with pLeontief is harder than Leontief or not when W 6= I. Apart from this, we identify
an interesting special case of p-Leontief utilities, which is tractable for Fisher
markets, which further gives us a non-trivial class of tractable exchange Leontief
markets.
Since Leontief function is also used in modeling production capabilities of
firms, for e.g. linear activity model [13], it would be interesting and useful to
apply p-Leontief function in those settings as well. In this paper, we define pLeontief as a concave function for modeling preferences of an agent, however
concavity can be easily discarded to define a more general function, which can
model all sorts of behavior in the production function of a firm, like initially it
is increasing returns to scale and then decreasing returns to scale etc. We expect
¶
This in fact works for any c > 0
more applications of piecewise Leontief in future.
Acknowledgement. The author is grateful to an anonymous referee for the
suggestion of uniformly scaled version of p-Leontief utility functions.
References
1. Arrow, K., Debreu, G.: Existence of an equilibrium for a competitive economy.
Econometrica 22, 265–290 (1954)
2. Brainard, W.C., Scarf, H.E.: How to compute equilibrium prices in 1891. Cowles
Foundation Discussion Paper (1270) (2000)
3. Chen, X., Deng, X., Teng, S.H.: Settling the complexity of computing two-player
Nash equilibria. Journal of the ACM 56(3) (2009)
4. Chen, X., Huang, L.S., Teng, S.H.: Market equilibria with hybrid linear-Leontief
utilities. Theor. Comput. Sci. 410(17), 1573–1580 (2009)
5. Codenotti, B., Saberi, A., Varadarajan, K., Ye, Y.: Leontief economies encode
two-player zero-sum games. In: SODA (2006)
6. Cottle, R., Pang, J., Stone, R.: The Linear Complementarity Problem. Academic
Press, Boston (1992)
7. Dang, C., Ye, Y., Zhu, Z.: An interior-point path-following algorithm for computing
a Leontief economy equilibrium. Comp. Opt. and Appl. 50(2), 223–236 (2011)
8. Daskalakis, C., Goldberg, P.W., Papadimitriou, C.H.: The complexity of computing
a Nash equilibrium. SIAM Journal on Computing 39(1), 195–259 (2009)
9. Dixon, P.B., Jorgenson, D.W.: Handbook of Computable General Equilibrium
Modeling SET, Vols. 1A and 1B. Elsevier ((Editors) 2013)
10. Eaves, B.C.: A finite algorithm for the linear exchange model. Journal of Mathematical Economics 3, 197–203 (1976)
11. Eisenberg, E.: Aggregation of utility functions. Management Sciences 7, 337–350
(1961)
12. Ellickson, B.: Competitive equilibrium: Theory and applications. Cambridge University Press (1994)
13. Gale, D.: Theory of Linear Economic Models. McGraw Hill, N.Y. (1960)
14. Lemke, C.E., J. T. Howson, J.: Equilibrium points of bimatrix games. SIAM J. on
Applied Mathematics 12(2), 413–423 (1964)
15. Mas-Colell, A., Whinston, M.D., Green, J.R.: Microeconomic Theory. Oxford University Press (1995)
16. Maxfield, R.R.: General equilibrium and the theory of directed graphs. Journal of
Mathematical Economics 27(1), 23–51 (1997)
17. Megiddo, N.: A note on the complexity of P-matrix LCP and computing an equilibrium (1988), http://theory.stanford.edu/~megiddo/pdf/plcp.pdf, IBM Research Report 6439
18. Shoven, J.B., Whalley, J.: Applying General Equilibrium. Cambridge University
Press (1992)
19. Vazirani, V.V., Yannakakis, M.: Market equilibrium under separable, piecewiselinear, concave utilities. Journal of ACM 58(3), 10:1–10:25 (2011)
´ ements d’´economie politique pure ou th´eorie de la richesse sociale
20. Walras, L.: El´
(Elements of Pure Economics, or the theory of social wealth). Lausanne, Paris
(1874), (1899, 4th ed.; 1926, rev ed., 1954, Engl. transl.)
21. Wilkens, C.A.: The complexity of models of international trade. In: WINE. pp.
328–339 (2009)
22. Ye, Y.: Exchange market equilibria with Leontief’s utility: Freedom of pricing leads
to rationality. Theor. Comput. Sci. 378(2), 134–142 (2007)
23. Zhu, Z., Dang, C., Ye, Y.: A FPTAS for computing a symmetric Leontief competitive economy equilibrium. Math. Program. 131(1-2), 113–129 (2012)